ruina.sun

cmake -version cmake version 3.17.3 CMake suite maintained and supported by Kitware (kitware.com/cmake).

Thanks so much! @ghost. However, how did you find the true reason is "num_output" not others? I have tried different lr list , but it does not work...

sampls_lst = os.listdir(data_path) sampls_lst_y = [] for i in sampls_lst: info = i.split('_') if len(info) >= 2: sampls_lst_y.append(info[1]) else: print(info) action_names = list(set(sampls_lst_y)) print(len(action_names), action_names)

> > @Hamiltonsjtu Hi, can you provide more detials? I also use the last get_embd.py, but still face this problem. I just put two different faces into one folder and...

> > > > @Hamiltonsjtu Hi, can you provide more detials? I also use the last get_embd.py, but still face this problem. I just put two different faces into one...

> @sunruina2 应该是作者的方法只reload了可训练的参数，有一些模型里frozen的参数和BN的参数就被忽略了会导致错误。另外之前这位答主用的import_meta_graph方法我也不是很了解，了解的话请赐教 @kscp123，做了4个对比测试供参考。主要原因是BN中的均值方差没有存在trainable_variables中，需要从global_variables取出来，再加入到trainable_variables里面，然后会得到正确结果了。和你执行的效果一样。(Four comparative tests have been done for reference. The main reason is that the mean and variance of BN do not exist in trainable_variables. It needs to...

另外，想请教一个问题：代码中emb的最终结果是：“原图embedding归一化” + “水平翻转原图embedding归一化”，然后再归一化一次。想请教一下，其中加上水平翻转embs的作用是什么呢？ （比较担心：如果原图是45度歪头或者侧脸的矩形的话，这样会不会使得最终结果会有“四只眼睛”或重影等问题呢？但实际运行来看即使不对称的face112原图，embds_arr和embds_f_arr两个结果很相似，是因为平移不变性对么？） embds_arr = embds_arr / np.linalg.norm(embds_arr, axis=1, keepdims=True) + embds_f_arr / np.linalg.norm(embds_f_arr, axis=1, keepdims=True) embds_arr = embds_arr / np.linalg.norm(embds_arr, axis=1, keepdims=True)

Is there a insightface pretrained model for this projects?